| Vose Software

Industry: Retail
Product: ModelRisk
Application: Analyzing Seasonal Trends and Demand Uncertainty


The Peak Has a 21% Chance of Arriving Outside Your Buy Window

Merchandising commits holiday-category stock to land in weeks 46–49 because the fitted seasonal index says demand peaks in week 48. The fitted index is a single curve. The real season is not. Simulate 40,000 plausible season shapes in ModelRisk and the peak week is a distribution — P10 week 46, median week 48, P90 week 50, a standard deviation of 1.5 weeks — and the probability the peak lands outside the planned buy window is 21%, almost all of it (16%) arriving late. Stock that lands four weeks before a late peak sits on the floor through the wrong fortnight.

This is not a forecasting-accuracy problem and it is not a promotion. It is the uncertainty in the shape of the recurring seasonal pattern itself — when the peak comes and how sharp it is — and the inventory risk of buying to the wrong week.

A season is a distribution of curves, not one curve

The deterministic approach fits one seasonal index per week from history and treats it as fixed. We instead generate a full season every trial, holding two properties that a real seasonal profile must satisfy: the 52 weekly indices average to 1.0 (an internally consistent profile, no negatives), and the weeks move together rather than independently.

  • A season-shift common factor (Normal, ±2.6 weeks at one sigma) slides the entire peak earlier or later — a warm autumn pushes the whole holiday curve back.
  • A peak-amplitude multiplier (LogNormal, CV 22%) controls how sharp the season is this year.
  • A width multiplier (LogNormal, CV 15%) decides between a sharp spike and a broad shoulder season.
  • A high-concentration Dirichlet adds gentle week-to-week wobble without breaking the sum-to-52 constraint.

Because the shift and amplitude are single per-year draws applied to the whole curve, the weeks are correlated by construction. The realised correlation between two off-peak weeks (week 40 and week 44) across trials is +0.84 — strongly positive, confirming the season moves as a coherent shape rather than 52 independent noises that would average away. That shared structure is the entire source of peak-timing risk.

Weekly seasonal index fan chart across the year

The fan chart shows the P10–P90 band of the seasonal index across all 52 weeks. The deterministic profile (dashed) is one smooth curve through the middle; the simulated band reveals both the lateral spread of the peak and how much higher a sharp year can spike. A single fitted index understates peak sharpness precisely because smoothing averages over the very years the buyer most needs to plan for.

Why a point estimate of "week 48" fails

The deterministic plan says one word — week 48 — and sizes the buy window around it. The simulation says the peak is a distribution with real mass on either side of the plan.

Distribution of the week in which demand peaks

Against the shaded weeks-46–49 buy window, the peak-week histogram shows P10 week 46, median week 48, P90 week 50. The 21% that falls outside splits 16% late, 5% early — the asymmetry matters because a late peak is the expensive failure mode: stock that arrived on schedule is already aging when demand finally crests. The deterministic "week 48" is the mode, but planning to the mode ignores the one-in-five seasons that miss it.

The cost of mis-timing the buy

Knowing the peak is uncertain only matters if it changes the buy. It does. If replenishment stock lands in week k and sells over the following four weeks, the share of peak-season demand actually captured depends sharply on k — averaged across all simulated seasons.

Share of peak demand captured versus the week stock lands

Landing in week 47 captures 95% of achievable peak-season demand; the planned window (weeks 46–49) captures 95% down to 83% as you slide later within it. But the curve is brutally asymmetric on the far side: land in week 50 and capture collapses to 59%; week 52 captures just 17%. Being two weeks early costs a few points; being four weeks late costs more than half the season. That asymmetry is the argument for buying to land slightly ahead of the modeled peak and holding flexible replenishment, rather than committing everything to a single week.

What moves the peak away from the plan

Tornado of drivers of peak-timing uncertainty

The season-shift common factor dominates peak-timing uncertainty (±2.5 weeks), well ahead of weather-onset timing and calendar drift. Amplitude and width multipliers reshape the peak's height but barely move its week. The lesson for data collection: a tighter read on what slides the whole season — weather onset, holiday-calendar position — narrows the timing risk far more than refining the peak's magnitude.

What the model changed

  • Buy-window placement. Knowing the capture curve is asymmetric, the team shifted target landing to week 47 and treated late arrival as the dominant risk, not a symmetric tolerance.
  • Flexible replenishment. With a 21% chance of an out-of-window peak, the plan reserved a fast-replenishment tranche instead of committing 100% of the season to a single landing week.
  • Right-sized peak inventory. The amplitude band (peak index P10 5.1, P90 6.6) replaced a single height assumption, so peak-week stock was sized to a credible interval rather than one number.

ModelRisk Functionality Used

  • Monte Carlo simulation of 40,000 full season shapes, turning a single fitted index into a peak-week distribution (P10 week 46, P90 week 50, std 1.5 weeks).
  • Dirichlet and shared common-factor modeling to generate seasonal profiles that average to 1.0 every trial while keeping weeks correlated (+0.84 between off-peak weeks) — the structure that creates genuine timing risk.
  • Parameter sweep of the stock-landing week, quantifying the asymmetric mis-timing cost: 95% capture at week 47 collapsing to 17% by week 52.
  • Tornado analysis isolating the season-shift common factor (±2.5 weeks) as the dominant driver of peak-timing risk, ahead of weather and calendar effects.

A fitted seasonal curve tells you when the peak usually is; only the distribution of curves tells you the one-in-five chance it arrives when your stock is already on the clearance rack.